Probability-Polarized Optimal Transport for Unsupervised Domain Adaptation
Authors: Yan Wang, Chuan-Xian Ren, Yi-Ming Zhai, You-Wei Luo, Hong Yan
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on several datasets validate the effectiveness and empirical efficiency of PPOT. We evaluate PPOT on three datasets. Office31 (Saenko et al. 2010) has 4,652 images from 3 domains with 31 classes, i.e., Amazon (A), Webcam (W), and Dslr (D); Image CLEF (Long et al. 2017b) has 3 domains with 12 classes, i.e., Caltech (C), Image Net (I) and Pascal (P); Domain Net (Peng et al. 2019) is a challenging large-scale dataset includes over 0.6 million images distributed across 345 classes from 6 different domains, i.e., Clipart (clp), Infograph (inf), Painting (pnt), Quickdraw (qdr), Real (rel) and Sketch (skt). The results on Office-31, Image CLEF and Domain Net are shown in Table 1. |
| Researcher Affiliation | Academia | Yan Wang1, Chuan-Xian Ren1*, Yi-Ming Zhai1, You-Wei Luo1, Hong Yan2 1School of Mathematics, Sun Yat-Sen University, China 2Department of Electrical Engineering, City University of Hong Kong, Hong Kong {wangy2277, zhaiym3, luoyw28}@mail2.sysu.edu.cn, rchuan@mail.sysu.edu.cn, h.yan@cityu.edu.hk |
| Pseudocode | Yes | Algorithm 1: PPOT Input: Maximum iteration Imax, parameters α and β; Output: OT plan γ ; 1: Initialize γ0 1 nsnt 1ns1T nt, ε0 1, i 0. 2: while i Imax and εi 1e 9 do 3: Compute Ωp(γi) and C + β Ωp(γi); 4: Obtain γ of Eq. (12) via Sinkhorn algorithm; 5: Search the optimal step αi via αi = arg min 0 α 1 f(γi + α γ) + g(γi + α γ), where γ = γ γi; 6: Update γi+1 = γi + αi γ; 7: Obtain OT distances L(γi+1) and L(γi) via Eq. (11); 8: εi+1 = |L(γi+1) L(γi)| 9: i i + 1. 10: end while |
| Open Source Code | No | The paper does not provide any explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | We evaluate PPOT on three datasets. Office31 (Saenko et al. 2010) has 4,652 images from 3 domains with 31 classes, i.e., Amazon (A), Webcam (W), and Dslr (D); Image CLEF (Long et al. 2017b) has 3 domains with 12 classes, i.e., Caltech (C), Image Net (I) and Pascal (P); Domain Net (Peng et al. 2019) is a challenging large-scale dataset includes over 0.6 million images distributed across 345 classes from 6 different domains, i.e., Clipart (clp), Infograph (inf), Painting (pnt), Quickdraw (qdr), Real (rel) and Sketch (skt). |
| Dataset Splits | No | The paper mentions 'train' and 'validation' in figures, and 'target domain' and 'source domain' for UDA, but does not specify explicit dataset split percentages, sample counts, or reference predefined splits with citations for reproducibility in the main text. |
| Hardware Specification | No | The paper states 'The implementation details and algorithm are presented in the appendix', but no specific hardware details (like GPU models, CPU types, or memory amounts) are provided within the main body of the paper. |
| Software Dependencies | No | The paper mentions 'The implementation details and algorithm are presented in the appendix' but does not specify any software dependencies with version numbers within the main text. |
| Experiment Setup | No | The paper states 'The implementation details and algorithm are presented in the appendix' and discusses the training principle, but concrete experimental setup details such as hyperparameter values (e.g., learning rate, batch size) are not provided in the main text. |